Runtime management

Runtime management is the boundary between the fuzzer and local processes used by stdio transports. It combines process lifecycle control, watchdog policy, signal delivery, and audit-safe cleanup.

Choose the right page

Need	Page
Understand process state, watchdog configuration, and platform signals	Process management
Start a target, tune timeouts, or troubleshoot cleanup	Process management guide
Understand bounded async/sync fuzz operation execution	Async executor
Trace how executors connect to mutators and reporting	Fuzz engine architecture

Runtime boundary

CLI / audit workflow
        |
Transport bootstrap ---- ProcessManager ---- ProcessRegistry
        |                                      |
        +------------------------------ ProcessWatchdog
                                               |
                                      SignalDispatcher

ProcessManager owns lifecycle operations. ProcessWatchdog reads the shared registry and applies configured hang policy. AsyncFuzzExecutor is deliberately separate: it bounds concurrent fuzz operations and does not supervise child processes.

Audit guidance

For local MCP security research:

1. Use an isolated working directory and dedicated test credentials.
2. Set --fs-root when generated tool arguments may contain filesystem paths.
3. Set --no-network unless the target's network behavior is in scope.
4. Use bounded --watchdog-* settings and preserve timeout/crash artifacts.
5. Run cleanup in a finally path when embedding the runtime API.

Runtime supervision is an operational control, not proof that a target is safe. Review runtime findings and server output as assessment evidence.

Source and tests

The implementation is under mcp_fuzzer.fuzz_engine.runtime. The maintained unit tests cover lifecycle transitions, watchdog scans, signals, registries, and cleanup behavior. Update the focused pages above when the runtime contract changes; do not add another copy of the executor API here.